ItemKNN with cosine similarity fails

[Djoels]

Dear all,

After running the itemKNN recommender using the following command:

item_recommendation --recommender=ItemKNN --recommender-options="k=1" \

     --test-ratio=0.0025 --random-seed=123 --measures="AUC,prec@5,prec@10,MAP,recall@5,recall@10,NDCG,MRR" \

     --training-file=data/MyMediaLite_ItemRecommender.tsv > MML_01_ItemKNN_k_1.log 2>&1

(I have done this for k=1, 5, 10, etc...)

The output is as follows:

Set random seed to 123.

loading_time 14.79

memory 171

training data: 257073 users, 238614 items, 11616612 events, sparsity 99.98106

test data:     23962 users, 21717 items, 29076 events, sparsity 99.99441

ItemKNN k=1 correlation=Cosine q=1 weighted=False alpha=0.5 (only for BidirectionalConditionalProbability) 

/../bin/item_recommendation: line 7: 116850 Killed                  mono --debug $DIR/../lib/mymedialite/item_recommendation.exe "$@"

There is no other mention of the requested evaluation measures (this is where the output stops).

More information: 

- there are 11000000 user-item interactions, they are in the right format (WRMF model works), there are roughly about 300000 users and 300000 items

- mymedialite has been recompiled with the large heaps flag set to true (which was a previous error)

Is there any idea about what could be going wrong here?

Regards,

Julien

[Zeno Gantner]

Hi Julien,

You just have too many users and items for MyMediaLite's kNN implementation.

The kNN implementations in MML do not scale well.

Cheers,

    Z.

--
You received this message because you are subscribed to the Google Groups "MyMediaLite" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mymedialite+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.